Dynamic Deep Learning Model Outperforms MELD 3.0 in Liver Transplant Prioritization

A clinical and computational research team has developed DynaMELD, a dynamic deep learning survival model designed to prioritize liver transplant waitlist candidates with greater precision and equity than existing national allocation scores. Trained on national registry data from 53,046 adult candidates across the United States, the neural network incorporates longitudinal rates of change in standard laboratory biomarkers alongside patient sex and specific diagnostic indications. In formal validation on a pre-pandemic evaluation cohort of 7,513 candidates, DynaMELD achieved a 90-day mortality concordance index of 0.827, representing a statistically significant 0.6 percent improvement over the current United States standard, MELD 3.0, which scored 0.822. Beyond overall accuracy, the model simultaneously corrected persistent structural disparities, improving discrimination and prioritization for women, patients with cholestatic liver disease, and shorter candidates who faced disadvantages under static allocation frameworks.

For over two decades, the Model for End-Stage Liver Disease and its successive iterations have served as the standard of care for deceased-donor liver allocation. Its current national version, MELD 3.0, estimates 90-day pretransplant mortality from five laboratory metrics—serum creatinine, bilirubin, international normalized ratio, sodium, and albumin—alongside sex. While MELD 3.0 partially mitigated sex disparities, persistent inequities remain for women, patients with cholestatic liver diseases such as primary biliary cholangitis and primary sclerosing cholangitis, and those experiencing acute-on-chronic liver failure. A major limitation of MELD is its static nature, relying on cross-sectional lab snapshots that fail to capture rapid clinical deterioration. Prior dynamic approaches like Delta-MELD collapsed multi-variable trajectories into a single composite score, obscuring variable-specific longitudinal signals that are most critical for identifying high-risk candidates within historically disadvantaged patient groups.

To address these challenges, the investigators structured DynaMELD as a neural network extension of the Cox proportional hazards framework underlying MELD 3.0. By maximizing the Cox partial likelihood directly, DynaMELD relaxes the rigid assumption of linear log-hazards, capturing complex nonlinear interactions among predictive biomarkers. The model incorporates serum albumin, creatinine, bilirubin, international normalized ratio, sodium, biological sex, and specific indicator variables for primary biliary cholangitis and primary sclerosing cholangitis. Furthermore, DynaMELD evaluates first-order rates of change representing biomarker velocity, computed from consecutive clinical observations normalized by elapsed time between measurements. After observing that patients with primary sclerosing cholangitis were insufficiently ranked relative to the broader waitlist, the researchers introduced an additive scalar bias term—yielding four DynaMELD points for the first-order rate model—by upweighting these cases in the Cox partial likelihood loss while preserving other network weights.

The derivation cohort was established using United States national registry records from the Scientific Registry of Transplant Recipients and Organ Procurement and Transplantation Network. Starting from 355,094 patients, the authors applied eleven sequential exclusion criteria to construct a clean adult cohort listed between January 1, 2016, and March 2, 2023. Key exclusions removed 259,013 candidates listed before 2016, 4,575 repeat transplants, 5,088 multi-organ listings, 3,106 acute liver failure cases, 317 HIV-positive recipients, 17,942 patients with hepatocellular carcinoma diagnoses, 638 listed under hepatocellular carcinoma exceptions, 7,048 with other exception points, 2,717 multi-listed or transferred cases, 1,319 pediatric or Status 1B candidates, and 285 pre-listing malignancies. The resulting cohort comprised 53,046 adult patients with 580,603 longitudinal observation instances. Patients listed from 2016 through 2018 were allocated to internal training, validation, and test sets using an 80/10/10 split, while 7,513 patients listed during 2019 formed an independent pre-COVID evaluation cohort.

In primary benchmark testing across 91,107 observations in the pre-COVID cohort, DynaMELD incorporating first-order rates of change achieved a 90-day concordance index of 0.827. Paired-sample bootstrap analyses demonstrated that DynaMELD significantly surpassed standard models, outperforming MELDNa at 0.816, the machine learning model OPOM at 0.817, and MELD 3.0 at 0.822, with every comparison showing statistical significance at p < 0.001. Integrating first-order biomarker rates of change provided a verified 0.3 percent improvement over static DynaMELD without velocity covariates, which scored 0.824. However, adding second-order acceleration terms yielded no additional predictive power, slightly decreasing concordance by 0.2 percent.

Beyond overall discrimination, the investigators evaluated fairness across clinical and demographic subgroups using the pooled group concordance index, which specifically measures ranking equity in survival data. In the pre-COVID test cohort, DynaMELD achieved statistically significant improvements over MELD 3.0 and MELDNa for female candidates, patients in the lowest height tertile measuring below 166 centimeters, and individuals diagnosed with primary biliary cholangitis or primary sclerosing cholangitis. For patients presenting with low-grade acute-on-chronic liver failure, DynaMELD achieved a pooled group concordance index of 0.853, significantly exceeding MELD 3.0 at 0.846. In more severe acute decompensation brackets, DynaMELD exhibited comparable concordance to MELD 3.0 among candidates with grade two and grade three acute-on-chronic liver failure, while significantly outperforming OPOM. These subgroup findings demonstrate that tracking individualized biomarker rates of change mitigates established biases without compromising predictive stability across high-acuity patient segments.

To determine how score adjustments influence organ allocation opportunities, the researchers performed a reclassification analysis on 1,340 patients from the pre-COVID test cohort who died on the waitlist while prioritized under MELDNa. DynaMELD reassigned 301 deceased candidates into higher risk quintiles compared to MELD 3.0, while shifting 178 into lower quintiles. This represented a net increase of 123 patients—or 9.2 percent of all waitlist deaths—who would have attained higher priority for donor organ offers prior to death. Furthermore, DynaMELD successfully reclassified a larger proportion of historically disadvantaged candidates into elevated risk quintiles relative to MELD 3.0 upon listing: 52.0 percent more patients with primary biliary cholangitis, 36.1 percent more with primary sclerosing cholangitis, 1.6 percent more female candidates, and 1.1 percent more candidates in the lower tertile of height, indicating meaningful structural improvements in pretransplant candidate prioritization.

To elucidate the internal decision-making of the deep learning model, the authors applied SHAP values across 5,000 sampled patient observations. The analysis demonstrated that standard MELD 3.0 laboratory variables carried the greatest overall predictive significance. Higher values of creatinine, bilirubin, and international normalized ratio strongly correlated with increased predicted mortality, whereas higher albumin and sodium levels lowered risk, matching physiological associations established in clinical literature. While rates of change showed modest average importance across the broader cohort, they exhibited pronounced outlier influence in acutely deteriorating individuals. Specifically, among 122 sampled instances where a rate-of-change variable generated a non-negligible SHAP value exceeding 0.2, 95.1 percent had received updated laboratory testing within the previous three days.

The generalizability of DynaMELD was evaluated through external validation on 699 patients listed at University Health Network in Toronto, Canada, between 2016 and 2022. Across the 455-patient external test cohort, DynaMELD achieved a 90-day concordance index of 0.826, which was statistically comparable to MELD 3.0 at 0.828. However, subgroup analyses demonstrated marked equity gains under DynaMELD across key vulnerable demographics. For female candidates, DynaMELD achieved a pooled group concordance index of 0.852 compared to 0.833 for MELD 3.0. For patients in the lowest height tertile measuring under 165 centimeters, DynaMELD scored 0.829 versus 0.810. Subgroup performance advantages were particularly substantial in cholestatic indications: DynaMELD reached 0.886 versus 0.834 for primary biliary cholangitis, and 0.881 versus 0.793 for primary sclerosing cholangitis.

To demonstrate prospective clinical deployment feasibility, the investigators integrated an interactive DynaMELD dashboard directly into University Health Network’s Epic electronic medical record system. The researchers conducted a simulated silent trial involving 382 patients waitlisted between June 2022 and September 2024, running model predictions in real time without altering bedside clinical care. Evaluating 30-day transplant probability via multivariable logistic regression that adjusted for donor supply and candidate demand factors demonstrated statistically significant increases in match probabilities under DynaMELD compared to MELD 3.0. In this silent trial cohort, DynaMELD reassigned 24 out of 105 waitlist deaths into higher risk quintiles compared to 11 moved lower, achieving a net 12.38 percent improvement in timely candidate prioritization. Furthermore, structured engagement with clinicians and patient partners confirmed the practical utility of the dashboard, while highlighting preferences for clear textual explanations over complex survival curves that could provoke unnecessary patient distress.

To evaluate systemic allocation dynamics, the researchers conducted policy simulations using the Liver Simulation Allocation Model, demonstrating that DynaMELD achieved statistically significant reductions in waitlist mortality for patients with primary biliary cholangitis and primary sclerosing cholangitis. Because deceased-donor liver allocation is inherently zero-sum, these survival gains were accompanied by modest mortality increases for other waitlist subgroups—a structural mathematical reality rather than an isolated model flaw. The authors also documented several study limitations. DynaMELD was designed specifically for MELD-based candidates, deliberately excluding patients listed with exception points such as hepatocellular carcinoma. In addition, MELD 3.0 retained superior concordance in alcoholic cirrhosis, and allocation simulations were constrained by the 2016 boundary of the simulation software. While normalizing biomarker rates of change by elapsed time introduces potential sensitivity to draw schedules, perturbation tests confirmed scores remained robust within three points. The authors concluded that continued international validation remains necessary before deploying deep learning allocation algorithms.

Subject of Research: Personalized dynamic deep learning model for equitable liver transplant waitlist prioritization

Article Title: Development and silent trial of DynaMELD, a Dynamic Model of End-Stage Liver Disease for personalized waitlist prioritization

Article References: Cooper, M. J., Khoroshchuk, D., Gao, X., Zhao, X., Wang, Y., Wang, X., Maksymowsky, P., Peralta, P., Bain, K., Azhie, A., Naghibzadeh, M., Holdsworth, S., Feld, J. J., Jaeckel, E., Brudno, M., Hirschfield, G., Krishnan, R. G., & Bhat, M. (2026). Development and silent trial of DynaMELD, a Dynamic Model of End-Stage Liver Disease for personalized waitlist prioritization. eClinicalMedicine, Article 104287. https://doi.org/10.1016/j.eclinm.2026.104287

Image Credits: AI Generated

DOI: 10.1016/j.eclinm.2026.104287

Keywords: DynaMELD, liver transplantation, waitlist prioritization, MELD 3.0, deep learning survival model, organ allocation, primary biliary cholangitis, primary sclerosing cholangitis, acute-on-chronic liver failure, sex disparities, concordance index, SHAP interpretability, electronic health records, silent trial, Scientific Registry of Transplant Recipients, University Health Network, time-varying biomarkers